Google opens Platform 37, a cutting-edge AI hub in London. See how DeepMind and GenAI startups are collaborating on the future of autonomous systems.
What Platform 37 is for
Platform 37 is Google’s London AI hub: a shared space where DeepMind research and GenAI startups can work on the same problems instead of in parallel silos. The focus named in the hub’s brief is autonomous systems—software and agents that plan, act, and adapt with limited human step-by-step control. That work is rarely pure research or pure product. It needs models that hold up under real constraints, evaluation that measures failure modes, and product teams that can ship safely.
A physical hub does not replace remote collaboration, but it changes the default. Engineers, researchers, and founders can sit through demos, debug edge cases on the same screen, and align on what “good enough to ship” means before a model leaves the lab. For autonomous systems, that shared context matters: the hard parts are often policy, tooling, and failure handling—not only the model weights.
How DeepMind and GenAI startups divide the work
DeepMind-style research tends to push on long-horizon reasoning, planning, and scientific method: strong baselines, careful ablations, and systems that improve under controlled experiments. GenAI startups usually optimize for product speed: integrating models into workflows, building UX around partial automation, and learning which tasks users will trust an agent to finish alone.
Collaboration works when each side keeps a clear role. Research can supply evaluation harnesses, safety checks, and techniques for multi-step control. Startups can supply real task distributions, latency budgets, and feedback from operators who live with the agent’s mistakes. Platform 37 is useful when that exchange is continuous—not a one-off talk or a press tour—so findings from production feed back into research priorities, and research constraints shape what products claim to automate.
Building autonomous systems without overpromising
Autonomous systems fail differently from chatbots. A wrong answer in a chat can be ignored; a wrong action in a workflow can cost money, time, or trust. Teams co-located around a hub like Platform 37 should treat autonomy as a ladder, not a switch:
- Start with draft-and-approve: the system proposes steps; a human confirms before anything external happens.
- Add scoped tools: allow only the APIs and data the task needs, with hard limits on side effects.
- Log every plan, tool call, and outcome so failures can be replayed and fixed.
- Expand autonomy only for tasks that meet clear success rates and reversible failure modes.
That ladder is product design as much as model design. Shared space between research and startups makes it easier to agree on the ladder’s rungs before marketing language outruns reliability.
Practical takeaways for builders watching London
You do not need a seat at Platform 37 to apply the same pattern. Pair long-horizon research questions with a narrow product surface. Define what the agent may do without asking, what always needs approval, and how you measure both task success and safe recovery. Prefer small, well-instrumented agents over broad “do anything” assistants until evaluation and ops catch up.
Platform 37’s value for the wider ecosystem is less about a single launch day and more about a repeatable loop: research improves planning and robustness; startups stress-test those ideas on real jobs; both sides tighten the path from prototype to dependable autonomy. If you are building in this space, design for that loop—shared metrics, shared failure reviews, and staged control—whether your partners sit across the table in London or across a video call.